Table 2 Training times (in hours) for each clinical NLP task and modeling paradigm. Values in parentheses indicate how many times longer the “Continue pre-training, fine-tune & predict” paradigm takes compared to the “Fine-tune & predict” paradigm. All training times were measured using a single NVIDIA RTX 4090 GPU. It is important to note that the reported times for the “Continue pre-training, fine-tune & predict” paradigm include the continuation of pre-training, which needs to be performed only once and can then be reused for multiple tasks

From: NLP modeling recommendations for restricted data availability in clinical settings